Berkeley and Nvidia release tools for agents to write and test robot programs

UC Berkeley and Nvidia have released open-source tools that let teams of AI coding agents generate, simulate and refine readable robot programs, The Robot Report reported. Their Graph-as-Policy framework divides work into skills such as perception, grasp planning and motion, which agents connect and rehearse in simulation. The accompanying VABAR benchmark evaluates six manipulation tasks, including gear meshing, syringe packing and cable routing, with standardized variations and difficulty levels. Generated Python code calls existing vendor modules for low-level control and can be converted to ROS 2. Bosch is evaluating the approach for industrial automation; the tools do not replace robot controllers.